2.1. Study population
We used data from the 1958 National Child Development Study (NCDS 58), a birth cohort study including all people born in Great Britain during one week in 1958 (n=18,555). Data collection on multiple aspects of the cohort members' lives, including health indicators, has been carried out from birth to age 60 by the Centre for Longitudinal Studies. The NCDS 58 has been described in detail elsewhere (25). In 2002–2004 when individuals were aged 44–45 years, information was collected through a biomedical survey (blood, saliva samples and anthropometric measurements) and a home-based clinical assessment, with data available for 9,377 individuals. The selection of this wave of data was relevant to conduct this study because it is the only wave in which both clinical and biomedical information were collected. Participants in this survey were found representative of the general cohort (26). A total of 2,334 participants were excluded from our analyses including pregnant women and those from whom blood was not obtained, as well as those with no data for the selected health reserve indicators and health outcomes, leaving a study sample of 7,043 participants (75% of the initial sample of the biomedical survey). The sample selection strategy is described in Additional Figure 1.
2.2. Health reserves
We first identified measurable items at 44/45y related to physical, psychosocial and physiological reserves to use them as indicators of deteriorating health reserves. No data on the deterioration of the cognitive reserve was available at this age.
Physical reserve: Chronic widespread Pain (CWP) is a common chronic pain measure used in several studies (27,28) and is defined according to the American College of Rheumatology Criteria for the Classification of Fibromyalgia (29): pain present for three months or longer, both above and below the waist; on both the left and right sides of the body; and in the axial skeleton. NCDS participants were asked ‘‘During the past month, have you had any ache or pain which has lasted for one day or longer?”. Those who answered positively were asked to indicate the location of their pain(s) on a four-view body manikin, and to indicate whether they had been aware of the pain for more than three months. Using this data, we constructed a binary variable identifying participants with CWP.
Psychosocial reserve: We used the revised CIS-r where scores of ≥12 indicate common mental disorders (21). In the NCDS 58, an abbreviated version of the CIS-r was used (sections enquiring after worry, obsessions, somatic symptoms, compulsions and physical health worries were omitted) (30). The domains that constitute this score in the NCDS cohort are: fatigue, concentration and forgetfulness, sleep problems, irritability, depression, suicidal ideation, anxiety, phobias and panic. Each domain provides a score from 0 to 4 (or 0 to 5 for suicidal ideation) based on the sum of their related items. The sum of these 14 domains resulted in an overall CIS-r score ranging from 0 to 33. We categorised this score into two groups “No mental health problem/Common mental health problem” according to the cut-off of 9 adapted from the abbreviated version of the CIS-r of the NCDS 58 (31).
Physiological reserve: We used data from the biomedical survey to construct an allostatic load score. Based on previous work carried out within the NCDS 58 (32), four physiological systems have been identified with the following 14 available and related biomarkers: the neuroendocrine system (salivary cortisol t1, salivary cortisol t1–t2); the immune and inflammatory system (insulin-like growth factor-1 (IGF1), C-reactive protein (CRP), fibrinogen, Immunoglobulin E (IgE)); the metabolic system (high-density lipoprotein (HDL), low-density lipoprotein (LDL), triglycerides, glycosylated hemoglobin (HbA1C)); the cardiovascular and respiratory systems: (systolic blood pressure (SBP), diastolic blood pressure (DBP), heart rate, peak expiratory flow). Using sex-specific quartiles, each biomarker was dichotomized into "high" (coded as 1) and "low" (coded as 0) risk. The sum of these 14 dichotomized biomarkers resulted in an overall AL score ranging from 0 to 14 where a higher score represented a high AL. This AL was also recoded into a 3 category variable based on tertiles in the sample, where a score of 0-2 was considered to be “low”, 3-4 as “middle ”, and 5-14 as “high” (33).
2.3. Global health score construction
We constructed the global health score variable using the selected indicators of deteriorating health reserves: CWP (2 categories), CIS-R (2 categories) and AL (3 categories). With regard to the theory previously developed, we supposed that each indicator of deteriorating health reserves has the same weight as the others. To ensure equal weighting between these variables, so that each variable varies from 0 to 1, we first reformatted AL to ensure that its number of categories (n=3) varied from 0 to 1. We rescaled its first category to 0, its second to 0.5 and its third to 1. Second, we summed AL, CWP and CIS-r. We finally classified our global health score into three groups “optimal [0]/medium[0.5-1]/impaired[1.5-3]”.
2.4. Health outcomes: Self-rated health and mortality
To validate our global health score we selected two health outcomes: mortality and self-rated health since health refers not only to being alive but also to living well (34).
All-cause mortality was derived from death certificates from the National Health Service Central Register recorded by the Centre for Longitudinal Studies. The mortality data most recently available to researchers provided information on date of death up to December 2016. The follow-up time was calculated from the date of blood collection to the date of death. Individuals with no death data but who had others data between the ages of 44/45 and 60 were classified as alive.
Self-rated health (SRH) was created using cohort members’ responses at 46y to the question “How would you describe your health generally?”. Responses were dichotomised into “excellent/good” versus “fair/poor/very poor”. In order to maximise the statistical power of our study sample, individuals with missing data at age 46 but who reported excellent/good perceived health at 50y and 55y (n= 580) were classified in the good self-rated health group at age 46 and those reporting fair/poor/very poor self-rated health at 50y and 55y (n= 138) were classified in the poor self-rated health group at age 46.
2.5. Statistical analysis
The content and construct validity of the global health score are ensured by its relevance to the theoretical framework developed in the introduction section. We explored criterion validity of the global health score by modelling the relationship between indicators of deteriorating health reserves, the global health score and other health outcomes (i.e. mortality and SRH). First, descriptive and bivariate statistics were carried-out to estimate the association of our selected indicators of deteriorating health reserves variables with self-rated health and mortality outcomes, using Chi square or logrank tests (35) as appropriate. Kaplan–Meier curves (36) were draw for participants’ indicators of deteriorating health reserves and participants’ global health score to validate the application conditions of the logrank tests. Second, associations between each of the indicators of deteriorating health reserves and the global health score with two health outcomes were modelled using logistic regressions for self-rated health and using multivariate Cox proportional regression (37) for mortality. Odds-ratio (OR) with 95% confidence intervals (CI) with hazard ratios (HR) and 95% CI were reported for the logistic regression models and Cox proportional regression respectively. We tested the proportional hazards assumption for each indicator of deteriorating health reserves and for the global health score. Both statistical testing using Schoenfeld residuals (38) and visual inspection (scatterplots of scaled Schoenfeld residuals vs. time) were performed and showed no violation of the proportional hazards assumption. Because perceived health and mortality differ between women and men (39,40), all models were adjusted for sex. We also estimated the sensitivities and specificities of the score in relation to mortality and perceived health.
2.6. Sensitivity analyses
We conducted sensitivity analyses to assess the robustness of the method. To ensure that the results observed were not biased by the grouping of the score, we tested an alternative grouping of the global health score distribution, combining the 0 and 0.5 categories and we performed the same multivariate analyses with the health outcomes (i.e. mortality and SRH). To ensure that the results observed were not biased by the variables constituting the score, we constructed other global health scores with a different categorisation of variables constituting it (global health score n°2 and n°3) or with a different choice of variables (global health score n°4) and we performed the same multivariate analyses with the health outcomes. A detailed description of the construction of the global health scores n°2, 3 and 4 and their respective variables is given in Additional File 1. To ensure that the results observed were not biased by the data of the sample, we generated two sub-samples using a negative control variable (or neutral variable) to perform multivariate analyses and assess the association between the global health score with the health outcomes. In the first sample, we selected all individuals who responded to the biomedical survey in odd-numbered months, and the second sample included all individuals who responded to the questionnaire in even-numbered months. We hypothesised that the trend of the results would be similar between the two sub-samples. We also carried out other sensitivity analyses to ensure that the estimates obtained of our health reserve indicators on self-rated health at age 46 did not differ at ages 50 and 55 (See Additional Figure 2).
All statistical analyses were performed using Stata ®v17 software (41).